Low Delay Audio Streaming for a 3D Audio Recording System
˙
Marzena Malczewska, Tomasz Zernicki
and Piotr Szczechowiak
Zylia Sp. z o. o.
Umultowska 85
Pozna´
n, Poland,
{marzena.malczewska, tomasz.zernicki, piotr.szczechowiak}@zylia.pl
Abstract
This paper presents a prototype of a 3D audio
recording system named AudioSense which uses
Wireless Acoustic Sensors to capture spatial audio.
The sound is recorded in real-time by microphones
embedded in each sensor device and streamed to
a Processing Unit for 3D audio compression. One
of the key problems in systems which stream audio
data is end-to-end latency. This paper is focused on
analyzing a set of chosen parameters of the Opus
codec in order to obtain the minimal delay. Experimental results on the prototype system have shown
that it is possible to achieve below 10ms of end-toend audio delay with the use of the Opus codec.
Keywords
audio streaming, opus codec, low latency streaming,
wireless sensor network, spatial audio
1
Introduction
The area of 3D audio and object based audio is
currently a hot research topic as evidenced by a
large number of research papers and new emerging standards such as MPEG-H 3D Audio. The
majority of the research eﬀorts in this area are
concentrated on the audio processing and rendering side. The problem of 3D audio recording
is getting less attention in the literature.
Channel-based method of spatial sound production assume that the number of microphones
used during the recording is directly proportional to the number of loud speakers used during sound rendering. This can lead to large
numbers of microphones in case of multiple audio channels recordings. In addition, proper
setting and tuning of microphones in the ﬁeld
can be a tedious task which requires many resources in terms of time and manpower. Current distributed recording systems use wired microphones, which makes it diﬃcult to deploy
and use the system in certain environments.
To solve the limitations of current spatial audio recording systems, the AudioSense system
is being developed. It introduces object-based
sound representation and wireless audio streaming. The system can be described as a Wireless
Acoustic Sensor Network (WASN) [Bertrand,
2011; Akyildiz et al., 2007]. In this system individual sound sources (objects) are extracted
from a sound mixture using sound source separation techniques (e.g Independent Component
Analysis [Comon, 1994]). Object based audio representation gives high ﬂexibility in terms
of sound rendering (easy rendering for headphones, stereo, 5.1, 7.1 systems) and enables
interactive manipulation of individual sound objects during playback.
The proposed AudioSense system has many
possible applications and can be used for both
indoor and outdoor audio recordings. The system can be used in teleconference applications
to add the possibility to identify speakers by
speech direction. It can be also used for wildlife
monitoring and live TV broadcasts from the
ﬁeld. The AudioSense technology has applications in surveillance systems where it can identify and track objects based on sound processing. The system can be also applied in the entertainment industry in case of movies, games
and virtual reality applications that require immersive 3D sound.
Realisation of the AudioSense system is a
challenging task that requires overcoming major challenges in such areas as audio streaming,
audio coding, sound sources separation and audio synchronisation. This paper is focusing on
designing a low delay audio streaming mechanism that meets the strict requirements of the
AudioSense system.
The AudioSense system consists of battery
operated devices with low processing capabilities. Therefore audio recording, coding and
streaming has to be performed with energy eﬃciency in mind. Live applications of the system
require also low latency audio streaming that
is reliable and allows simultaneous streaming
of data from multiple devices over the wireless
medium.
In order to minimise the end-to-end delay
it is necessary to optimise the audio recording process, use a low latency audio codec and
streaming method. This paper shows the design of the system that tries to achieve this goal.
It presents the technologies and design choices
made to implement a low delay audio streaming
system on oﬀ-the shelf embedded devices. The
results achieved during the performance evaluation of the system show that it is possible
to achieve a low end-to-end delay of 10ms for
wireless audio streaming within the AudioSense
system.
The rest of the paper is organised as follows.
Section 2 presents the related work in the area
of spatial audio recording systems. Section 3 describes the architecture of the AudioSense system together with hardware and software components implemented to build the ﬁrst prototype of the system. The results achieved during
the performance evaluation phase are presented
in Section 4. Finally Section 5 concludes the
paper and describes the lessons learnt from implementing a low delay audio streaming system.
2
Related work
Spatial audio recording systems are gaining on
popularity with the introduction of 3D audio
systems and technologies that can reproduce
truly immersive sound. Majority of existing
systems for spatial audio recording use wired
microphones [Gallo et al., 2007] to capture the
virtual sound stage. This fact limits drastically
the mobility of such systems and increases signiﬁcantly their deployment time.
In the literature one can ﬁnd also wireless systems for distributed audio recording like [Taysi
et al., 2010] and [Pham et al., 2014]. The main
problem with such systems is that the wireless
sensor network devices are equipped with low
quality microphones, ampliﬁers and A/D converters due to the low cost and high energy efﬁciency of the system. Sounds recorded with
such systems have insuﬃcient quality for many
audio applications.
One of the systems that tries to overcome
the problems of low cost WASNs is WiLMA
[Sch¨orkhuber et al., 2014]. The system introduces a wireless microphone array that offers high quality audio recording and processing. WiLMA enables connection of up to 4
professional microphones to each sensor module
and provides wireless synchronisation for audio
recordings. Similar approach to system design
is presented also in [Mennill et al., 2012] where a
distributed microphone array system is used for
environmental monitoring and animals recording. This system is also based on battery operated sensors and uses GPS for accurate synchronisation of the recordings.
One of the limitations of spatial audio recording systems presented in [Mennill et al., 2012] is
that the system does not oﬀer continuous realtime wireless audio streaming. All the recordings are stored on local ﬂash memory of the
sensor devices. The AudioSense system takes
the next step in spatial audio recording systems
by providing low delay wireless streaming capabilities and audio representation in the objectbased format. These features open up the door
for a whole new range of audio applications that
can be realised with the use of the AudioSense
system.
3
System overview
The proposed architecture of the AudioSense
system is presented in Figure 1. From the functional side the system can be divided into two
parts. The ﬁrst part consists of Acoustic Sensors that form a wireless network responsible
for audio recording. The second part includes
an embedded device which performs 3D audio
processing. Each device in the wireless sensor
network has one or several microphones, A/D
converter and performs initial audio compression. Compressed audio signals are transmitted
through the Gateway to the Processing Unit.
The Gateway serves as an interface between the
wireless and the wired part of the system. After reception of the audio signals the Processing Unit performs aggregation of the individual
streams.
Each of the streams is decoded and synchronised with each other. In the next step the process of sound sources separation is performed
to generate individual audio objects [Sala¨
un et
al., 2014], [Ozerov et al., 2012]. These objects
are then used in the process of 3D audio coding
(e.g. MPEG-H 3D Audio [ISO/IEC WD 230083, 2014]. Finally the encoded audio is transmitted over the Internet to the client side where the
sound rendering is performed.
3.1 Hardware components
From the hardware perspective the Acoustic
Sensor prototype consists of a Beaglebone Black
[Coley, 2014] with an Audio Cape board [BeagleBoard, 2012] and a dedicated microphone
Figure 1: Architecture of the AudioSense system.
board designed in-house. Beaglebone Black is
a low-power single board computer based on
1GHz ARM Cortex A8 CPU. The board provides only one 12-bit analog-to-digital converter
which is not suﬃcient for any professional audio applications. Hence the usage of an Audio Cape (6 channels of up to 96 kHz sampling
at 24 bit) is required to improve the quality of
the recorded sound. For the microphone board
a pair of Monacor MCE-4000 electret omnidirectional microphones is selected due to high
signal to noise ratio and very good sensitivity.
Each microphone is connected to a low noise
operational ampliﬁer - MCP6021. The ampliﬁed acoustic signal is passed on to the Audio
Cape where analog to digital conversion is executed. Next, the digital data in one of available
formats (e. g. S16LE) is sent to the Beaglebone Black. Each acoustic sensor is equipped
also with a wireless interface compatible with
the IEEE 802.11 a,b,g,n standards. The ﬁrst
version of the prototypical Acoustic Sensor is
presented in Fig. 2.
Figure 2: First version of the Acoustic Sensor
prototype.
3.2
Software components
The prototype of Audio Sensor is running Debian Jessie Linux with kernel version 3.8.13.
In order to implement audio processing on the
device, the Gstreamer framework is utilised.
Sound capturing, coding and streaming are
all implemented within a single Gstreamer
v1.4 pipeline. Figure 3 illustrates Gstreamer
pipelines implemented on both the Acoustic
Sensor and the 3D Audio Processing Unit.
The pipeline on the Acoustic Sensor side is
responsible for capturing audio samples using
the ALSA plugin and encoding them with the
Opus [Valin et al., 2013] encoder. Next, every
packet is encapsulated in the RTP packet and
sent via UDP to the Processing Unit.
On the Processing Unit side each received
packet is processed by the depayloader and
Opus decoder. This processing is performed for
each stream independently. Next step is the
separation plugin, which gets n streams and,
after performing the process of sound sources
separation, generates m audio objects. The processed data is passed on to the multiqueue and
then interleaved to form a multichannel wave
ﬁle. For this purpose a new Gstreamer module is implemented called WavNChEnc. The
stream generated by the module is then passed
to the MPEG-H 3D Audio codec which generates a single .mp4 ﬁle.
To measure time of encoding, the measurements points were set right before and after
Opus encoder. Respectively, on the Processing Unit the points were set before and after
the Opus decoder. Streaming time was measured with the measurement points set just before UDP transceiver in Acoustic Sensor and
right after RTP depayloader in the Processing
Unit.
One of the key aspects in networked audio
systems is audio synchronisation. In order to
provide synchronisation of the recorded audio
streams, a separate synchronisation module is
implemented. The synchronisation method applied is a hybrid approach based on reference
broadcast that uses the ideas presented in [Elson et al., 2002] and [Budnikov et al., 2004].
Using this hybrid synchronisation method it is
possible to achieve a synchronisation error of
around 200µs.
which includes:
• Audio encoding time - time needed to encode one whole buﬀer of data by the Opus
encoder. Such measurements were executed for diﬀerent codec parameters which
have the highest impact on the encoding
time (e.g. bitrate, complexity, frame-size).
• Audio decoding time - measurement of decoding time for the same set of parameters
as in the case of audio encoding.
• Audio transmission time - packets latency
measurement when streaming wirelessly
over Wi-Fi (IEEE 802.11n).
For the Audio streaming tests the Opus parameters were constant while the network setup
was diﬀerent in each experiment. The system
was tested with several Acoustic Sensors in the
network. In each of the cases the distance between the Acoustic Sensors and the 3D Audio
Processing Unit was diﬀerent to test the system
in diﬀerent working conditions. In addition to
latency tests the experiments included also CPU
usage measurements for Opus encoding and decoding.
Impact of selected parameters of the Opus
codec on the quality of sound was not the subject of our tests. Several tests were performed
in the past and are well described in [Hoene et
al., 2011].
Figure 3: Gstreamer pipelines implemented on
the Acoustic Sensor and the 3D Audio Processing Unit side.
4
Performance evaluation
4.1 Test scenarios
The performance of the designed 3D Audio
recording system was evaluated using several
test scenarios. The main goal of the experiments was to adjust and optimise hardware and
software components of the system to achieve
minimal streaming delay for diﬀerent audio bitrates and network setups. Audio streaming latency was measured in an end-to-end manner
4.2 Results
This subsection presents experimental results
achieved by measuring the end-to-end audio
streaming delay in the AudioSense system. The
ﬁrst set of tests was performed to measure the
encoding delay of the Opus codec in order to
ﬁnd the optimal codec parameters that enable
minimal processing latency. Three parameters
of the codec were identiﬁed as possible candidates for processing delay optimisation:
• Complexity - is deﬁned as a trade-oﬀ
between processing complexity and quality/bitrate. This parameter is selected using an integer from 0 to 10, where 0 is the
lowest complexity and 10 is the highest. In
the experiments ﬁxed values of 0, 3, 6 and
10 were used to check what is the inﬂuence
of complexity on the processing delay.
• Frame size - Opus has ﬁxed frame durations of 2.5, 5, 10, 20, 40, and 60 ms. Increase in the frame duration has inﬂuence
on coding eﬃciency improvement but the
gain becomes small for frame sizes above
20 ms.
• Bitrate - Opus supports diﬀerent bitrates in
the range between 6 kbit/s and 510 kbit/s.
Higher bitrate results in higher quality audio and lower latency in packets delivery at
the cost of increased bandwidth.
Time of coding
bitrate: 128 kbps
30
25
Time [ms]
20
2.5
5
10
20
40
60
15
10
5
0
0
3
6
10
Complexity
Figure 4: Opus encoding time for diﬀerent values of complexity and frame-size.
The diﬀerence is especially visible for higher
frame durations of 40 and 60 ms where the latency is 3 to 6 times higher than in the case of
smaller frame sizes. Therefore the best values of
frame size in case of the AudioSense system are
below 20ms where the encoding delay is smaller
than 10ms. Surprisingly the frame size of 2.5ms
is providing a similar encoding delay as in the
case when the frame duration is set to 20ms. In
terms of complexity the optimal value is 3 with
frame size of 10 ms.
The inﬂuence of diﬀerent audio bitrates on
the Opus encoding delay is illustrated in Figure 5. The complexity parameter in all cases
is ﬁxed at 0. The experiments are performed
for ﬁve audio bitrates (64, 96, 128, 256 and 320
kbit/s) and the same frame size values as in the
previous test. The graph shows that signiﬁcant
increase in processing time is visible for larger
values of frame size (40 and 60 ms). It is evident that the bitrate change has much smaller
eﬀect on encoding time than the change of the
complexity parameter. For frame size values below 20ms the change of bitrate has very small
eﬀect on the encoding delay - only 1ms increase
when changing the bitrate from 64kbit/s to 320
kbit/s. This experiment shows once again that
the frame size of 10ms provides the optimal setting in terms of Opus encoding latency.
Time of coding
complexity: 0
Time of decoding
15
bitrate: 128 kbps
9
8
7
10
6
Time [ms]
Time [ms]
2.5
5
10
20
40
60
2.5
5
10
20
40
60
5
4
5
3
2
1
0
64
96
128
256
320
Bitrate [kbps]
0
0
3
6
10
Complexity
Figure 5: Opus encoding time for diﬀerent values of bitrate and frame-size.
Figure 4 shows Opus encoding delay for different complexity and frame size settings (diﬀerent colors on the ﬁgure correspond to diﬀerent
frame sizes). For all measurements the bitrate
remained constant at the level of 128 kbit/s. It
is clearly visible that the average encoding latency increases with higher complexity values.
Figure 6: Opus decoding time for diﬀerent values of complexity and frame-size.
Opus decoding latency is tested in a similar
manner as in the case of encoding. Figure 6
shows the impact of the complexity parameter
on the decoding time for diﬀerent frame duration values. The audio bitrate is ﬁxed at 128
kbit/s. As can be seen in Figure 6 the com-
Time of decoding
CPU usage
complexity: 0
60
8
7
55
6
50
2.5
5
10
20
40
60
4
3
CPU [%]
Time [ms]
5
2.5
5
10
20
40
60
45
40
2
35
1
0
30
64
96
128
256
320
64
96
128
256
320
Bitrate [kbps]
Bitrate [kbps]
Figure 7: Opus decoding time for diﬀerent values of bitrate and frame-size.
Figure 8: CPU usage for diﬀerent values of bitrate and frame-size.
plexity parameter has a small impact on the
overall audio decoding time. For smaller values of frame size (10ms and below) the decoding
time remains the same for diﬀerent complexity
values. The diﬀerence is visible only in case of
larger frame size values (20, 40 and 60ms) where
the decoding delay can increase or decrease by
around 1ms with the change of codec complexity.
Figure 7 presents Opus decoding times with
respect to diﬀerent audio bitrates. In all cases
the complexity parameter is set to 0. It is
clearly visible that audio bitrate has very small
impact on the decoding time. The main parameter that has the biggest inﬂuence on decoding time is frame duration. From the point of
view of Opus decoding, the best performance in
terms of execution time can be achieved for the
smallest possible values of frame size: 2.5 and
5ms.
The AudioSense system consists of battery
operated sensor devices therefore the power consumption during codec operation is an important parameter that can limit the total operation time of the system. Figure 8 presents the
CPU usage on the Acoustic Sensor while performing coding and decoding using the Opus
codec. The measurements are performed for six
diﬀerent audio bitrates and six frame durations.
In all cases the CPU operation remains between
35% and 57%. Highest CPU usage is recorded
for the smallest frame size value (2.5ms). For
all frame sizes between 10ms and 60ms the CPU
usage stays on the same levels. The inﬂuence of
audio bitrate on CPU processing is not signiﬁ-
cant as changing the bitrate from 64 kbit/s to
320 kbit/s increases the CPU usage by 7% on
average. The complexity parameter of the Opus
codec has a stronger inﬂuence on the CPU processing than audio bitrate change. Changing
the complexity from 0 to 3 increases the CPU
usage by 10% on average. Switching from 3 to 6
adds another 10% of CPU processing. It is recommended to set the complexity on 3 or lower
in order to keep the CPU usage below the level
of 50%. From the energy eﬃciency point of view
the optimal frame size is equal to 10ms.
3.50
3.25
Latency of streaming - distance ca. 1 m
1 x AS
2 x AS
3 x AS
3.00
2.75
2.50
Time [ms]
2.25
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0.00
0
500
1000
Measurement no.
1500
2000
Figure 9: Streaming delay measurements with
a distance of 1m between devices.
Audio encoding and decoding adds a significant delay in the audio processing pipeline of
the AudioSense system. The third factor that
adds an additional delay is audio streaming
over the wireless channel. In order to mea-
3.50
3.25
Latency of streaming - distance ca. 6 m
1 x AS
2 x AS
3 x AS
3.00
2.75
2.50
Time [ms]
2.25
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0.00
0
500
1000
Measurement no.
1500
2000
Figure 10: Streaming delay measurements with
a distance of 6m between devices.
3.50
3.25
Latency of streaming - distance ca. 10 m
1 x AS
2 x AS
3 x AS
3.00
2.75
2.50
Time [ms]
2.25
2.00
1.75
1.50
1.25
1.00
0.75
0.50
0.25
0.00
0
500
1000
Measurement no.
1500
2000
Figure 11: Streaming delay measurements with
a distance of 10m between devices.
sure streaming latency over Wi-Fi several experiments are performed using the prototype AudioSense system.
All the tests are made with the same parameters of the Opus codec: sampling rate 48kHz,
bitrate 128 kbit/s, complexity 0, frame size
10ms. Figure 9 presents the ﬁrst set of experiments where the network consists of one, two
or three Acoustic Sensors (AS). In all cases the
distance between the 3D Audio Processing Unit
and Acoustic Sensors is equal to 1m. The measurements are taken over 2000 audio samples.
The streaming is performed under Line of Sight
(LOS) conditions using the 802.11 n mode. It
is clearly visible in Figure 9 that the streaming
delay is the lowest (around 70µs) when there is
only one Acoustic Sensor in the network. Addition of the second sensor that sends simultaneously audio data to the processing unit in-
creases signiﬁcantly the overall packets delivery
time to around 1ms on average. The network
with three Acoustic Sensors increases the delay
even further to around 1.6ms.
Figures 10 and 11 show the results of the
same experiment as above but under diﬀerent
network conditions. The distance between the
devices is increased to 6m and 10m respectively.
The streaming is performed under Non Line of
Sight (NLOS) conditions. The results demonstrate that the increase in distance between devices has small inﬂuence on the average audio
streaming delay. The average delay for packet
reception remains at the same levels in all three
sets of tests. The main diﬀerence can be noticed
in the jitter levels which are much higher when
using the system in NLOS conditions.
5
Conclusions
This paper presents the architecture of the AudioSense system which is designed to record and
process spatial audio. All the hardware and
software components of the prototype implementation of the system are described in detail. The result of the work is a wireless acoustic sensor network capable of distributed sound
recording in an object-based audio format.
The paper focuses also on the development
of a low delay audio streaming technique which
meets the strict requirements of the AudioSense
system. For this purpose the Gstreamer framework is utilised together with the Opus codec.
The optimal working parameters for the codec
are selected through experimental evaluation
and the end-to-end delay is measured for different setups of the wireless network. The results demonstrate that it is possible to achieve
an average delay below 10ms for coding, transmission and decoding of the audio signal in a
wireless system of several Acoustic Sensors.
For the future work it would be interesting
to test the system on a larger scale with parallel transmissions from many Acoustic Sensors.
The capacity of the system and transmission delay can be further optimised by utilising wireless streaming in the 802.11 ac standard. For
the needs of sound sources separation it will be
beneﬁcial to apply a hardware based synchronisation method which would limit the synchronisation error to several µs.
6
Acknowledgements
This work was supported by National Centre for
Research and Development (NCBiR), Poland,
”Leader” programme.
References
Adapteva, 2014. Parallella Reference Manual
13.11.25.
Ian F Akyildiz, Tommaso Melodia, and
Kaushik R Chowdury. 2007. Wireless multimedia sensor networks: A survey. Wireless
Communications, IEEE, 14(6):32–39.
BeagleBoard, 2012. BeagleBone Audio Cape
Revision A1 System Reference Manual, October.
Alexander Bertrand. 2011. Applications and
trends in wireless acoustic sensor networks:
a signal processing perspective. In Communications and Vehicular Technology in the
Benelux (SCVT), 2011 18th IEEE Symposium on, pages 1–6. IEEE.
D. Budnikov, I. Chikalov, S. Egorychev,
I. Kozintsev, and R. Lienhart. 2004. Providing common i/o clock for wireless distributed
platforms. In Acoustics, Speech, and Signal Processing, 2004. Proceedings. (ICASSP
’04). IEEE International Conference on, volume 3, pages iii–909–12 vol.3, May.
Gerald Coley, 2014. BeagleBone Black System Reference Manual - Revision B, January.
Pierre Comon. 1994. Independent component
analysis, a new concept? Signal processing,
36(3):287–314.
Jeremy Elson, Lewis Girod, and Deborah Estrin. 2002. Fine-grained network time synchronization using reference broadcasts. In
Proceedings of the 5th Symposium on Operating Systems Design and implementation,
OSDI ’02, pages 147–163, New York, NY,
USA. ACM.
Emmanuel Gallo, Nicolas Tsingos, and Guillaume Lemaitre. 2007. 3d-audio matting,
post-editing and re-rendering from ﬁeld
recordings. EURASIP: Journal on Advances
in Signal Processing. Special issue on Spatial
Sound and Virtual Acoustics.
Christian Hoene, Jean-Marc Valin, Koen Vos,
and Jan Skoglund. 2011. Summary of opus
listening test results. Technical report, IETF.
ISO/IEC WD 23008-3. 2014. Information
technology - High eﬃciency coding and media
delivery in heterogeneous environments - Part
3: 3D audio. Technical report, International
Organization for Standardization / International Electrotechnical Commission, International Telecommunications Union – Telecommunication, January.
Daniel J Mennill, Matthew Battiston,
David R Wilson, Jennifer R Foote, and
Stephanie M Doucet. 2012. Field test of an
aﬀordable, portable, wireless microphone array for spatial monitoring of animal ecology
and behaviour. Methods in Ecology and Evolution, 3(4):704–712.
Alexey Ozerov, Emmanuel Vincent, and
Fr´ed´eric Bimbot. 2012. A General Flexible
Framework for the Handling of Prior Information in Audio Source Separation. IEEE
Transactions on Audio, Speech and Language
Processing, 20(4):1118 – 1133, May. 16.
Congduc Pham, Philippe Cousin, and Arnaud Carer. 2014. Real-time on-demand
multi-hop audio streaming with low-resource
sensor motes. In Local Computer Networks
Workshops (LCN Workshops), 2014 IEEE
39th Conference on, pages 539–543. IEEE.
Yann Sala¨
un, Emmanuel Vincent, Nancy
Bertin, Nathan Souvira`a-Labastie, Xabier
Jaureguiberry, Dung T. Tran, and Fr´ed´eric
Bimbot. 2014. The Flexible Audio Source
Separation Toolbox Version 2.0. May.
Christian Sch¨orkhuber, Markus Zaunschirm,
and IO-hannes Zm¨olnig. 2014. Wilmawireless largescale microphone array. In
Linux Audio Conference, volume 2014.
Z. Cihan Taysi, M. Amac Guvensan, and
Tommaso Melodia. 2010. Tinyears: Spying
on house appliances with audio sensor nodes.
In Proceedings of the second ACM Workshop
on Embedded Sensing Systems for EnergyEﬃciency in Buildings, BuildSys ’10, pages
31–36, New York, NY, USA. Association for
Computing Machinery.
Jean-Marc Valin, Gregory Maxwell, Timothy B Terriberry, and Koen Vos. 2013. Highquality, low-delay music coding in the opus
codec. In Audio Engineering Society (AES)
Convention no. 135. Audio Engineering Society.